Rank-profile evaluation

Recall@K for each rank profile against a sampled validation set. Methodology lives in demos/visual-rag/eval/; rerun with uv run python evaluate.py --samples 200 --seed 42.

Latest run: · samples

profileR@1R@5R@10R@20
bm25
colpali
hybrid

The cells render as — until the first eval run is committed. The harness in the repo regenerates this table on demand against the live corpus — see the README in /eval/.